Collider bias creates a false relationship by looking at the wrong slice of the data.
It is the mirror image of confounding: confounders need adjusting for, while colliders need leaving alone.
A collider is a common effect, such as fame, which both talent and good looks feed into.
In the full population, talent and looks may be unrelated.
Among the famous, though, anyone short of one must have plenty of the other to have become famous.
That produces a negative link between talent and looks that exists only inside the selected group.
The same thing happens whenever data is filtered on an outcome that both variables affect.
Hospital data is a classic case, since 2 unrelated illnesses can look linked among people admitted for either.
Studies of only successful companies, top students or published research are prone to the same bias.
Adjusting for a collider in a model, not just filtering on it, has the same effect.
Some biases blamed on confounding are really collider bias, which is why drawing the graph matters.
Adjusting for a consequence of a collider opens the false link too, so the danger extends further down the graph.
The fix is to keep colliders and their consequences out of the adjustment and to avoid samples selected on them.
In practice, asking whether a filter is based on something both variables affect catches most cases.
The usual mistake is adjusting for every available variable in the hope of removing bias, which can add collider bias.